Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Instability thing...

Author: Richard Pijl

Date: 15:50:22 09/18/04

Go up one level in this thread


On September 18, 2004 at 11:28:45, Sune Fischer wrote:

>On September 18, 2004 at 11:09:26, Robert Hyatt wrote:
>
>>On September 18, 2004 at 04:05:38, Sune Fischer wrote:
>>
>>>
>>>I decided to try out the triangular PV thing Bob
>>>speaks so highly of, to see if it improves move ordering...
>>>
>>>I was careful to terminate the PV on all exact scores - of course.
>>>Still I was getting illegal moves in the PV.
>>>
>>>It turned out to be a hash/nullmove problem.
>>>
>>>See, in the hash I adjust the window down on UPPER bounds, ie.
>>>something like:
>>> if (flag==UPPER && score<beta)
>>>    beta = score;
>>
>>Are you doing PVS?  If so you won't do this a dozen times during a long search,
>>most likely.  But I do this in Crafty as well, and it doesn't _ever_ get illegal
>>moves in the PV.
>
>I haven't looked but I guess you are doing something equivalent to what Richard
>suggested with beta = score+1.
>

It may be less important when doing fail-hard. Chances that the upperbound
equals the real score is lower with fail-hard.
Richard.

>Small but important difference. :)
>
>>>The idea is that if we know the score is in the interval [x,y]
>>>then there is no reason to search with a window of [x,y+z].
>>>It's a common trick I believe.
>>>
>>>Unfortunately, the search will occasionally fail high on the new beta, in my
>>>case it was a nullmove that failed-high.
>>>
>>>That's clearly contradictive with what the hash just told us so
>>>it doesn't actually make any sense, but apparently it happens anyway
>>>(instability?).
>>>
>>>Now the score at the previous node might be an exact value (or a fail high later
>>>on), and the PV will now be concatenated from an old search branch.
>>>
>>
>>That can't happen.  You don't update the PV on a fail high, ever.  Only if score
>>is > alpha and < beta...
>
>The PV is not updated on that node either, it's updated above at the parent node
>because here the FH-score will fall within the window.
>
>And of course that will result in a random clipping of the PV.
>
>This doesn't seem possible to prevent a 100%, because how do we guarantee that
>the new search will produce the same score as the old one when the hash has been
>changed meanwhile?
>
>-S.



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.